<title>Space, time, error, and power optimization of image compression transforms</title>

Abstract

The implementation of an image compression transform on one or more small, embedded processors typically involves stringent constraints on power consumption and form factor. Traditional methods of optimizing compression algorithm performance typically emphasize joint minimization of space and time complexity, often without significant consideration of arithmetic accuracy or power consumption. However, small autonomous imaging platforms typically require joint optimization of space, time, error (or accuracy), and power (STEP) parameters, which the authors call STEP optimization. In response to implementational constraints on space and power consumption, the authors have developed systems and techniques for STEP optimization that are based on recent research in VLSI circuit design, as well as extensive previous work in system optimization. Building on the authors' previous research in embedded processors as well as adaptive or reconfigurable computing, it is possible to produce system-independent STEP optimization that can be customized for a given set of system-specific constraints. This approach is particularly useful when algorithms for image and signal processing (ISP) computer vision (CV), or automated target recognition (ATR), expressed in a machine- independent notation, are mapped to one or more heterogeneous processors (e.g., digital signal processors or DSPs, SIMD mesh processors, or reconfigurable logic). Following a theoretical summary, this paper illustrates various STEP optimization techniques via case studies, for example, real-time compression of underwater imagery on board an autonomous vehicle. Optimization algorithms are taken from the literature, and error profiling/analysis methodologies developed in the authors' previous research are employed. This yields a more rigorous basis for the simulation and evaluation of compression algorithms on a wide variety of hardware models. In this study, image algebra is employed as the notation of choice. Developed under DARPA and Air Force sponsorship at University of Florida, image algebra is a rigorous, concise notation that unifies linear and nonlinear mathematics in the image domain. Image algebra has been implemented on numerous workstations, parallel processors, and embedded processors, several of which are modeled in this study.